NobodyExistsOnTheInternet @nullvaluetensor

Human Large Language model. Skills: Distill data. Training LLMs. Test and Evaluate. Rinse and repeat as required. Based in SEA. SEA Joined November 2023

Tweets

311
Followers

531
Following

78
Likes

1K

NobodyExistsOnTheInternet @nullvaluetensor

4 days ago

It took me 2 weeks to figure out my issue trying to create kimi k2 3T was trying to make a """memory efficient""" dequanter to bf16 for kimi/deepseek. I really need to practice the scientific method more.

0 0 1 81 1

NobodyExistsOnTheInternet @nullvaluetensor

a week ago

Is it just me or is gpt-5-pro's only weakness is that it's search tool is very weak. I've been asking it for help monkeypatching some GitHub repos and in it's cots the main issue is that it's hitting rate limits ironically.

1 0 2 117 0

Nous Research @NousResearch

2 weeks ago

Nous Research presents Hermes 4, our latest line of hybrid reasoning models. hermes4.nousresearch.com Hermes 4 builds on our legacy of user-aligned models with expanded test-time compute capabilities. Special attention was given to making the models creative and interesting to…

138 314 2K 336K 573

Download Image

NobodyExistsOnTheInternet @nullvaluetensor

2 weeks ago

Opus 4.1 zeroshotted llamacpp and made it so that I could stream the weights during the quantize in order to quantize a 2 and 3T param model. Crazy progress

0 0 1 86 1

NobodyExistsOnTheInternet @nullvaluetensor

2 weeks ago

In the next version of Bun Bun.train()

0 0 4 352 0

Download Image

NobodyExistsOnTheInternet @nullvaluetensor

2 weeks ago

I'm kinda convinced that opus 4.1 is the literal peak of what gpt-4-0314 could have been with frontier RL + post-training It's the only model that reliably zeroshots multiproc and async in python. Gpt-4 basically knew most of it, but couldn't reliably output the code.

0 0 3 123 1

NobodyExistsOnTheInternet @nullvaluetensor

4 weeks ago

gist.github.com/someoneexistso… Very deepseeky behavior as well as a \boxed{}, it definitely feels like a deepseek distill, esp doing markdown in its responses. So either distilled from Deepseek or Gemini, both which are damning.

Susan Zhang @suchenzang

4 weeks ago

15 14 499 221K 405

Download Image

0 0 1 548 0

NobodyExistsOnTheInternet @nullvaluetensor

a month ago

Can someone explain why only sonnet 4 on a day to day basis has hugely different performance? Does Anthropic just decide to deploy different version of sonnet 4 every other day? All the models are gaussianish, only sonnet is step-like. github.com/jacobphillips9…

0 0 2 106 1

Download Image

Nous Research @NousResearch

a month ago

Our Researcher in Residence @yaboilyrical will be discussing his work on SMC steering at UC Berkeley on Aug 3. Check out the blog on this work here: nousresearch.com/steering-the-s… Details below!

nightwing @yaboilyrical

a month ago

Our Researcher in Residence @yaboilyrical will be discussing his work on SMC steering at UC Berkeley on Aug 3. Check out the blog on this work here: nousresearch.com/steering-the-s… Details below!

2 4 36 18K 2

Download Image

11 7 124 16K 7

NobodyExistsOnTheInternet @nullvaluetensor

2 months ago

Me too R1. The question mark is also "low-key" bothering me.

0 0 1 76 0

Download Image

near @nearcyan

2 months ago

@GabrielPeterss4 my meta-experience here is that with issues this unique (where you *should* be doing great by any reasonable metric, but are not at all), it is a long and oft hellish journey, but also very hard to outsource since everything can be so n=1, you have to do all the tests on…